def deconv(data,
output_shape,
scope,
shape,
stride=[1, 2, 2, 1],
padding='SAME',
wd=0.0,
FLOAT16=False,
reuse=None):
with tf.variable_scope(scope, 'DeConv', [data], reuse=reuse):
STDdev = 1 / tf.sqrt(
shape[0] * shape[1] * shape[2] / 2) #Xavier/2 initialization
kernel = _variable_with_weight_decay('weights',
shape=shape,
stddev=STDdev,
wd=wd,
FLOAT16=FLOAT16)
deconv = tf.nn.conv2d_transpose(data,
kernel,
output_shape,
stride,
padding=padding)
biases = _variable_on_cpu(
'biases', [shape[2]],
tf.constant_initializer(0.00001)) #positive biases
pre_activation = tf.nn.bias_add(deconv, biases)
sm._activation_summary(pre_activation)
return pre_activation
def fclayer(data,batch_size,hidden,scope,wd=0.0,FLOAT16=False,reuse=None):
with tf.variable_scope(scope, 'fc',[data],reuse=reuse):
# Move everything into depth so we can perform a single matrix multiply.
reshape = tf.reshape(data, [batch_size,-1])
dim = reshape.get_shape()[1].value
weights = _variable_with_weight_decay('weights', shape=[dim, hidden],
stddev=0.04, wd=wd,FLOAT16=FLOAT16)
biases = _variable_on_cpu('biases', [hidden], tf.constant_initializer(0.00001))
pre_activation = tf.matmul(reshape, weights) + biases
sm._activation_summary(pre_activation)
return pre_activation
def dilated_conv(data,scope,shape,rate=1,padding='SAME',wd=0.0,FLOAT16=False,reuse=None):
with tf.variable_scope(scope, 'Dilated_Conv', [data], reuse=reuse):
STDdev=1/tf.sqrt(shape[0]*shape[1]*shape[2]/2) #Xavier/2 initialization
kernel = _variable_with_weight_decay('weights',
shape=shape,
stddev=STDdev,
wd=wd,FLOAT16=FLOAT16)
conv = tf.nn.atrous_conv2d(data, kernel, rate, padding=padding)
biases = _variable_on_cpu('biases', [shape[3]], tf.constant_initializer(0.0001))#positive biases
pre_activation = tf.nn.bias_add(conv, biases)
sm._activation_summary(pre_activation)
return pre_activation
def batch_norm_layer(x, train_phase, scope_bn, reuse=None):
with tf.variable_scope(scope_bn, [x], reuse=reuse):
z = batch_norm(x,
decay=0.999,
fused=False,
center=True,
scale=True,
is_training=train_phase,
reuse=reuse,
trainable=True,
scope=scope_bn,
updates_collections=None,
variables_collections=[
"batch_norm_non_trainable_variables_collection"
])
sm._activation_summary(z)
return z
def leakyReLU(x,leak, scope, reuse=None):
with tf.variable_scope(scope, 'leakyReLU', [x], reuse=reuse):
leakyrelu= tf.nn.leaky_relu(x,alpha=leak,name=None)
sm._activation_summary(leakyrelu)
return leakyrelu
def ReLU(x,scope, reuse=None):
with tf.variable_scope(scope, 'ReLU', [x], reuse=reuse):
relu= tf.nn.relu(x)
sm._activation_summary(relu)
return relu